Discrete and Continuous Submodular Bandits with Full Bandit Feedback
Wednesday, September 27, 2023
In this presentation, we will delve into our recent breakthroughs in the realms of discrete and continuous submodular bandits. For discrete submodular sequential decision-making, we introduce a versatile framework designed to transform discrete offline approximation algorithms into sublinear α-regret methods that exclusively rely on bandit feedback. Remarkably, this framework demands no more than resilience to errors in function evaluation from the offline algorithms. Even more intriguingly, the adaptation process does not necessitate explicit knowledge of the inner workings of the offline approximation algorithm; it can be seamlessly utilized as a black-box subroutine. Furthermore, we address the issue of feedback delays in this context. Shifting our focus to continuous submodular optimization, we will unveil a unified approach meticulously crafted for the maximization of continuous DR-submodular functions. This approach proves to be highly versatile, accommodating a wide spectrum of settings and oracle access types. It encompasses a Frank-Wolfe type offline algorithm tailored for both monotone and non-monotone functions, even under diverse constraints on the convex set. We explore scenarios where the oracle provides either gradient information or just function values, and where the oracle access is deterministic or stochastic. We also elucidate the precise number of required oracle accesses for each case. Notably, our method for the stochastic function value-based oracle provides the first-ever regret bounds with bandit feedback for stochastic DR-submodular functions.
Post Talk Link: Click Here
Passcode: 77U#VNUj
Speaker/s
Vaneet Aggarwal received the B.Tech. degree from the Indian Institute of Technology Kanpur, Kanpur, India, in 2005 and the MA and PhD degrees from Princeton University, Princeton, NJ, USA, in 2007 and 2010, respectively, all in Electrical Engineering. He is currently a Full Professor at Purdue University, West Lafayette, IN 47907, USA, where he has been since 2015. Prior to that, he was with AT&T Labs Research (2010-2014). He has been a Visiting Professor at KAUST, Saudi Arabia (2022-2023), VAJRA Adjunct Professor at IISc Bangalore (2018-2019), and Adjunct Assistant Professor at Columbia University (2013-2014) Dr. Aggarwal received the Princeton University's Porter Ogden Jacobus Honorific Fellowship in 2009 and Purdue University's Most Impactful Faculty Innovator award in 2020. In addition, he received IEEE Jack Neubauer Memorial Award in 2017, IEEE Infocom Workshop Best paper award in 2018, and Neurips Workshop Best paper award in 2021. He was an Associate Editor for the IEEE Transactions on Green Communications and Networking (2017-2020) and IEEE Transactions on Communications (2017-2022). He has served/is serving as Senior Program Committee member for AAAI 2023, IJCAI 2023, AAAI 2024, and ICLR 2024. He is currently serving on the Editorial Board of the IEEE/ACM Transactions on Networking (2019-current), and is co-Editor-in-Chief of the ACM Journal on Transportation Systems (2022-current). His current research interests include reinforcement learning, federated learning, quantum machine learning, and applications of machine learning.
Related
Polygenic Score Modeling to Investigate Genotype-Phenotype Associations
Read More
Trustworthy Machine Learning: Transparency, Collaboration, and Evaluation
Read More